(744c) A NEW Framework for Online Optimization of Recombinant Protein Production in FED-Batch Fermentation Processes

Online optimization of protein production is a challenging problem in the pharmaceutical industry. A new framework for online optimization of protein production in fed-batch fermentation processes is proposed in this study. This framework consists of trajectory planning and nonlinear model predictive control (NMPC) to realize online optimization of fed-batch fermentation. To avoid the effects of measurement noise and model-process mismatch, the proposed framework also includes dual Unscented Kalman Filter (UKF) to estimate both process states and model parameters, simultaneously. An online optimization of tissue-plasminogen activator (tPA) protein expression in mammalian cell culture is studied by using the proposed framework. In this case study, a modified Monod type kinetic model is applied to precisely predict the states of bioprocess including the nutrients consumption rates, cell numbers in all state of the cells (live, dead. lysed), and product formation rates. With the fed-batch mammalian cell culture model, the trajectory states are determined by solving a nonlinear optimization problem for maximizing the protein productivity. The calculated trajectory states are taken as reference states in NMPC. To avoid control degradation due to model uncertainties and perturbation during the real-time fermentation, a dual estimation strategy based on Unscented Kalman Filter is developed to estimate the actual process states and model parameters at each sampling time. To study the influence of the updated model parameters on the protein productivity in the fed-batch fermentation, variance-based global sensitivity analysis is performed to estimate the importance of the model parameters. The updated model parameters with their total-effect indices are used to evaluate the extent of the process-model mismatch. If the value of inner products of the vector of model parameters and the vector of their corresponding total-effect indices is over a threshold, the trajectory planning is executed again to update the trajectory states of the fermentation. Through the trajectory planning and NMPC algorithm, the optimal feeding rates and concentrations are determined for maximizing the productivity of tPA-I and tPA-II proteins during fermentation. The simulation result of protein production shows 12.64% improvement compared with the batch fermentation. The preliminary results show that under conditions of noisy measurement and structured plant-model mismatch, the dual estimation algorithm can precisely estimate the actual states.